A Neural Architecture for Pattern Recognition Insensitive to Translation, Scale, and Line Thickness

Abstract

In this paper, an architecture for pattern recognition system using neural networks is presented which is insensitive to translation, scale and line thickness. The system consists of a preprocessing network and a trainable multilayered network. An input pattern is first normalized in position and size through a pattern standardizing network. A feature extraction network is next used to detect line features which are an activity pattern in the OSC (Orientation Specificity Cell) layer. The line features are also used as an input pattern to the multilayered network which is then trained to recognize the pattern. Computer simulations of a numeral recognition task show the effectiveness of the system.